Google Vertex AI Model Registry and Versioning

Google released support for versioning models earlier this year. However, the official documentation on how to use it is still very limited. This article should help you shed light on the Vertex AI SDK and how to use it to version your models.

Until recently, most of us were confronted with an overly cluttered list of models.

Page over pages with the same model trained on different days. This inelegance was caused because we didn’t have a way to upload different versions of a specific model. Each version we train, for example, as part of a Vertex AI Pipeline, was uploaded as a separate model.

Luckily this has changed, as Google introduced a bunch of additional features to manage our model versions properly.

How to use model versioning?

You can use the model versioning obviously via Cloud Console (UI). Yet, most models are built automatically, and we likely need to interact programmatically. Luckily we can use the SDK and API for that.

API

As engineers, we first check the API documentation. As we all know, Google is API first. And indeed, the v1beta1 contains an API endpoint for uploading different versions of a model.